The author compared four acceleration schemes for MOSS-Image SFT on an RTX 3060 12GB; the Kijai 4step LoRA shortened the original ~83-second 50-step sampling to 4.3–4.4 seconds, about 19x.
Akashiba, an operations member of note's #ハロウィンアート2026 event, was drawn in by an image generation prompt sent by organizer KITAcore while organizing submitted magazines, and couldn't resist trying some doodling. As of writing, the event had received 133 submissions, with the deadline on the 31st.
The author used ChatGPT to generate AI illustrations, asking for emphasis on particle feel, opacity and perspective, referencing the light effect of car headlights in the rain; ChatGPT provided multiple options at once and explained the common elements. The final choice was a forest scene: foreground and background vegetation naturally distributed, light intensity and direction supplementing the layers, dark-area depth enhancing the sense of presence, and leaves with low noise blending into the background without stealing focus from the subject. The author believes AI illustrations are better described in terms of Lighting and Writing to convey atmosphere rather than stated plainly.
The organizer of the Halloween Art 2026 project introduced 18 selected works from submissions, covering AI illustrations, music, video and articles in various forms. The works include music made with Suno, GIF animations, PVC figure-style character prompts, and creative approaches that weave AI dialogue into narrative. The project has no judging or ranking, and the full works are published on a dedicated exhibition website.
A user installed ComfyUI on a Ryzen AI Max 395+ with 64GB of memory, using LTX-2.5 to generate 10-second clips at 720p, 24fps, aiming to assemble an approximately 5-minute Ghostbusters Slimer video, but each of the 10 results had problems—first growing legs, then voices, then green or white buttocks appearing. He plans to stitch about 300 segments, with each render taking about 10 minutes, and is seeking a more stable model or prompt solution, with plans to later upgrade to 1080p and then upscale to 4K.
Krea released an Illustrator plugin that can generate images within an Illustrator document, then use Image Trace to convert them into editable vector paths. The plugin supports sending selections or artboards back to Krea for modification and placing them back in place; Krea Agent can draw shapes, typeset text, and organize layers and artboards within the document.
Krea released an After Effects plugin that generates motion concepts, background plates and coherent shot footage within the panel, with results entering the current composition as standard media assets. The plugin supports After Effects 2025 and above, macOS 12 or Windows 10 and above; generation runs on Krea servers and requires account login and internet connection; the panel has built-in Krea Realtime preview, and the free plan includes 100 compute units per day.
Why it matters: Krea brings image generation into the After Effects panel, letting motion designers generate images within compositions and use them directly as assets.
Shri Labs open-sourced Shri Tools (MIT license), which can produce lip-synced cartoon short films and music videos offline on an RTX 3050 4 GB VRAM laptop. The pipeline uses FLUX.2 Klein 4B to generate consistent character footage, MiniMax H3 (GGUF) to drive speech and lip sync, and a vision model plus Whisper to self-check and redo item by item, supporting English plus 11 Indian languages.
The default Edit node of Fibo 2.0 Edit supports only 1MP out of the box; a user used OpenCode and the free LLM MiniMax M3.1 Flash to generate a custom ComfyUI node in about 10 minutes in one go, raising the limit to 2MP. The node has some issues with detail and element transitions when exceeding 1MP, but is generally usable.
A beginner using Easy Diffusion on a 1050ti found that generation speed only had two levels, 8s/it and 60s/it, with almost no intermediate values, and no clear correlation with prompt length or LoRA.
A user reported that MINIMAX H3 is praised by the community as an excellent editing model, but their own attempts only produced low-quality or distorted human anatomy results. The user asked the community for help, hoping someone would share H3 image editing workflows as well as image generation workflows.
The user trained a detail and texture LoRA "RAW Upscale MOSS" for MOSS-Image using LoRA Dataset Studio, trained on MOSS-Image-Base for 2000 steps, rank 16 / alpha 16, 1024 resolution, finally exported as FP16 SafeTensors, and has been made freely available for download on Civitai.
A ComfyUI user upgraded from version 3.x to 4.2, and multiple previously well-functioning workflows developed blurry images, broken details such as eyes and lips, and backgrounds and textures blurring together, with LoRA and parameters unchanged. The user uses krea2 and found that LoRA strength above 1.0 seems to cause problems, but lowering it still couldn't restore the previous clear results, and is asking about the cause and how to roll back to 3.x.
DirCast uses a single streaming video language model to simultaneously perform directing and commentary for live esports broadcasts: it scans all player camera feeds at low resolution every second, a hold-or-switch directing module decides whether to cut the camera, and then it reads the selected camera feed at high resolution before speaking commentary.
ComfyUI workflow templates released v0.11.94, installable via pip install comfyui-workflow-templates==0.11.94. This version adds three new templates: Mirelo SFX, VEED, and HeyGen Voice 1.0 Turbo.
Higgsfield's Prompt Team compared Seedance 2.5 and Veo 3.1 across five criteria: prompt responsiveness, motion stability, lighting, character consistency, and editing control.
Nastya Yangar's AI satirical short film "Grow Baby" imagines a world where babies can be bought like commodities and raised at home in 21 days. In the film, a young woman's expectations of parenthood are shaped by advertising and social pressure; everyday conversations, promotional slogans, and matter-of-fact retail transactions make this premise seem ordinary, revealing the tension between "choosing a child" and "accepting a child."
Alexandre Tissier released the Absolut Mexico AI concept ad, using hybrid AI imagery to connect ancient Mexican civilizations, the colonial encounter, mariachi, Día de los Muertos, and lucha libre. The film revolves around "Who says Mexico only has one soul?" It first sets up expectations of tequila, then reveals vodka in a final twist.
Bear Atelier released the AI gothic fantasy short film "The Estate Sale," about an auction of curiosities from a deceased collector, where a final unlisted lot reveals the extraordinary creature he took in and cherished. The film uses AI-generated images, video, and audio, with the final shots completed using Higgsfield. With lavish velvet-and-gold settings and an eccentric auctioneer, it turns a eerie premise into warm dark humor about the difference between monetary value and true value, paying homage to Guillermo del Toro's compassion for monsters.
The Vathography author released the ComfyUI-Marigold-V2-Tiled-Depth-Vathography node, which can estimate depth maps from high-resolution photos. The node is free, its output is commercially usable, and an example workflow is included. The author says the node should run on an 8 GB GPU plus about 24 GB of system memory, pending confirmation, and is soliciting hands-on feedback from different GPUs.
ComfyUI launched FLUX 3 Max, a latency-optimized version of FLUX 3 Video that can generate 10-second 720/768p clips with native audio in about 12 seconds. The official release also provides links to two workflows, image-to-video and text-to-video, and lists use cases such as comparing multiple candidates, previewing motion while tuning parameters, batch-producing ad variants, and keyframe-guided storyboards.
Why it matters: ComfyUI integrates FLUX 3 Max, offering a latency-optimized version that produces 10-second 720/768p clips with native audio in about 12 seconds, along with entry points for image-to-video and text-to-video workflows.
The MOSS-Image community released a 2GB offline character and style catalog containing 33,937 characters and 221 styles, each with one-click copying of generation prompts. In the same period, a LoRA for removing the posed look, a night flash street-photography LoRA, a workaround workflow for image-to-image based on Qwen visual descriptions, a low-VRAM LoRA training tutorial, and an Apache-licensed Spectrum accelerator appeared. The MOSS team's technical report is freely available on Arxiv.
The author tested Bria's FIBO 2.0 Edit Turbo and found it can also perform text-to-image when no input image is provided: 4 steps, guidance_scale=1.0, only 4 transformer forward passes per render, whereas the Dev version requires 50 steps plus CFG.
After updating ComfyUI to run Moss Image, the user has to wait a long time on every startup, restart, or browser refresh; the startup task has completed but the interface is blank, and another refresh is needed before it finally loads.
MIPCOM 2026 was held in Cannes, focusing on the transformation of the audiovisual industry, the integration of traditional TV and AI, the rise of micro-dramas, and strategic partnerships with platforms such as YouTube, Meta, and TikTok. It also announced that the 2027 edition will take place from October 18 to 21. MIPJUNIOR, as a prelude, will be held on October 10 and 11, with Disney as the main focus, discussing children's content strategies. During the event, companies like Inter Medya showcased new projects such as micro-dramas.
OpenFill beta.2 was released, a free local generative fill plugin for GIMP 3 running on ComfyUI, supporting generative fill, object removal, generative expand, and text-to-image, with the selection acting as the mask.
The author released the ComfyUI-AudioCPP project, aiming to unify audio.cpp's capabilities—TTS, ASR, voice conversion, music generation, audio enhancement, and separation—into ComfyUI, covering 120+ audio model families.
Kling's official account teased the next-generation video generation model with #Kling4, without giving specific parameters in the post. The Morphic release information it cited says Kling 4.0 supports native 30-second generation, up to 10 keyframes, 15 references, on-screen text consistency, and multiple languages, and offers 4K and 10-bit HDR, with a 10% discount for generations on Morphic before November 20.
Why it matters: Kling officially teases Kling 4; cited information indicates it supports native 30-second generation and 4K 10-bit HDR, which can be used to gauge its long-shot output capability.
The Shanghai Stock Exchange issued a regulatory warning to Nanwei Software and then-board secretary Zhang Yuan for inaccurate, incomplete, and inadequately risk-disclosed information about its AI short drama and intelligent computing center businesses disclosed on an interactive platform and in media interviews. The company had said its AI short dramas had annual output of over 100 and could reach more than 300 in the future, and projected large-scale revenue by 2027, but confirmed that the business currently generates no revenue.
Kling announces that the KLING 4.0 API is officially open, allowing developers to integrate Kling 4.0 into their own products and workflows. The original text does not disclose specific capability parameters, pricing, or integration details.
The creator locally calls the Seedance 2.0/2.5 API in ComfyUI to generate clips, and combines them with MiniMax H3 shots, Suno music, and a self-trained Z-Image LoRA character to synthesize a 1-minute-40-second AI music video. This post is a showcase of the finished work and does not include a reusable workflow file.
AIArtEngine releases v8.3.0, with a full changelog covering the changes from v8.2.0 to v8.3.0. The original text does not disclose specific features, parameters, or availability details.
10/21/26 generative media news roundup: The Verge reports on the proliferation of celebrity AI face-swapping, Variety says the sci-fi thriller "Overstake" received a generative AI reconstruction and was picked up by Cappu Films at the Busan market. Deadline focuses on how Asian directors are integrating Kling AI into their workflows, and openmoss.ai releases MOSS-Image, a text-to-image model supporting native layout control.
The Youku costume drama "Jia Jin Chai" broke 10,000 in platform popularity, setting a new record for the year, with its Yunhe market share rising to 26%, and the Lighthouse daily market climbing to 290 million, up more than 50 million compared with before the drama aired. The drama broke through with its twin black lotus villainess setup and dual-strong CP rivalry, but controversies coexist, such as the Jun Dai character being superficial and logical loopholes in the script. The article uses this to discuss whether the villainess narrative and the short-drama-ification dividend are nearly exhausted after multiple works in the same lane underperformed during the summer season.
AI Canvas v0.10.13 supports double-clicking reference capsules to view prompts, storyboards, character libraries, and other content in full screen; video generation can read the current shot's framing, camera movement, dialogue, and actual duration based on stable shot references, and reordering shots will not mix them up.
The article proposes an approach of first writing the storyboard in text and generating images only after confirmation: break the planned scenes into shots, write out the presentation order and the overall/medium/close-up framing for each, then have ChatGPT output a storyboard plan according to a unified project. Using a fictional umbrella shop 20-second, 4-shot video as an example, it requires the AI not to generate images before confirmation, and to revise line by line from four points: narrative coherence, visual consistency, shot function, and production volume.
The author sets up a MiniMax H3 environment in ComfyUI and tests pablodawson's MiniMax-H3-360-Orbit-LoRA, used to make still images generate an orbiting rotation effect.
TapNow (Tamar AI)'s international AI film hackathon Arena Peak Asia was held from October 16 to 18, 2026, in Shanghai, Tokyo, and Seoul; about 160 teams were selected from roughly 3,200 registrations to complete short films within 32 hours.
Runway announces that Kling 4 is available for unlimited use on its platform, and users who newly subscribe to the Max or Team plans can get 3 days of unlimited quota. All plans also enjoy a 25% discount on Kling 4.